TWI509376B - Automatic fault detection and classification in a plasma processing system and methods thereof - Google Patents

Description

Automatic fault detection and classification in plasma processing system and method thereof

The present invention relates to automatic fault detection and classification in plasma processing systems and methods therefor.

The present application claims the priority of the co-owned provisional patent application "Automatic Fault Detection and Classification in Plasma Processing Systems", U.S. Patent Application Serial No. 61/222,098, attorney Docket No. P2006P/LMRX-P182P1 The application was filed on June 30, 2009 by the inventor, the entire disclosure of which is incorporated herein by reference.

Plasma processing systems have long been used to process substrates such as semiconductor wafers and flat panels. The plasma processing system is used to perform processes such as deposition, etching, cleaning, and the like.

For example, in plasma processing systems used to produce semiconductor components, plasma processing systems are highly expected to produce electronic components with the highest possible yield and lowest cost of ownership. To achieve high yields and reduce tool downtime (which contributes to higher cost of ownership), it is critical to quickly detect and classify faults to minimize damage to the wafer and/or to plasma processing system components. For example, failure conditions can result from failure of chamber components, wear of chamber components, improperly installed chamber components, and/or the need to clean, maintain, and/or replace more than one plasma processing system. Any other case of the subsystem.

Modern plasma processing systems use a variety of sensors to monitor various process parameters such as light emission, voltage, current, pressure, temperature, and more. Data monitoring performed by each sensor outputs data at a rate of up to hundreds of samples per second or more. Given the large number of sensors involved, modern plasma processing systems generate large amounts of sensor data for a particular processing wafer. If the sensor data analysis is performed manually, it is usually impossible to accurately detect and/or classify the fault condition from a large amount of sensor data in time. If the fault condition cannot be detected in time, further processing can cause damage to more than one wafer and/or to the chamber components. Even after the plasma treatment is suspended, it takes a lot of time to screen a large amount of sensor data to confirm the failure that has occurred to help repair the fault.

Manual fault detection and analysis also requires highly skilled engineers to screen a very large amount of data. These highly skilled engineers are not only short of shortages but also expensive to hire, both of which increase the cost of ownership of the machine owners. And the procedures for manual fault detection and analysis are also prone to errors.

In the past, attempts have been made to automatically detect fault conditions and analyze sensor data to classify faults. The results of these efforts have achieved varying degrees of success in the production environment and in the market. Engineers continually search for ways to detect fault conditions faster and classify faults more accurately. This application is an improved method and apparatus for automatically detecting fault conditions and classifying fault conditions automatically and in a timely manner.

An embodiment of the invention is directed to a method of automatically detecting a fault condition and classifying a fault condition during substrate processing. The method includes collecting processing data by a set of sensors during substrate processing. The method also includes transmitting the processing data to a fault detection/classification component. The method further includes performing a data operation of the processing data by the fault detection/classification component. The method further includes comparing the processing data with a plurality of fault models stored in a fault database, wherein each fault model of the plurality of fault models represents a set of data depicting a particular fault condition feature. Each fault model includes at least one fault trace, a fault boundary, and a set of principal component analysis (PCA) parameters.

The above summary is only one of the many embodiments of the invention disclosed herein, and is not intended to limit the scope of the invention. These and other features of the present invention will be described in detail in the following embodiments of the present invention.

The invention will now be described in detail with reference to a number of embodiments as illustrated in the accompanying drawings. In the following description, numerous specific details are set forth in the description. However, it will be apparent to those skilled in the art that the present invention may be practiced without some or all of the specific details. In other instances, well known process steps and/or structures have not been described in detail in order to avoid unnecessarily obscuring the present invention.

The following description includes various embodiments of the methods and techniques. Should be kept in mind The invention may also encompass manufacturing products, including computer readable media stored by computer readable instructions for performing embodiments of the present technology. By way of example, computer readable media includes semiconductor, magnetic, photomagnetic, optical, or other forms of computer readable media for storing computer readable code. In addition, the present invention also encompasses an apparatus for carrying out embodiments of the present invention. Such devices include circuitry (dedicated and/or programmable) to perform tasks related to embodiments of the present invention. Examples of such devices include general purpose computers and/or dedicated computing devices written with appropriate programming, and include combinations of computer/computing devices and special/programmable circuits suitable for various tasks related to embodiments of the present invention.

Embodiments of the present invention are directed to highly automated, time-sensitive, and robust methods of fault detection and classification that can be directed to any sensor data set associated with more than one wafer in a test.

For purposes of discussion, FIG. 1 presents an example of a lower electrode subsystem 100 that includes a lower electrode 102, an outer cover ring 104, and a top cover ring 106. Wafer 108 is also presented. The top cover ring 106 is shown as partially worn, representing an example of a fault condition type that would affect the plasma and negatively alter the process results. In the production environment, timely detection of the fault condition depicted in Figure 1 and timely and accurate classification of the fault is highly anticipated in relation to the worn top cover ring, which can prevent the subsequent processing of the substrate and/or the plasma Handle damage to other components of the system and quickly restore the system after repair/maintenance.

2 is a logic block diagram showing various sub-components 200 of a plasma processing chamber in accordance with an embodiment of the present invention, which can automatically and timely detect fault conditions and automatically and timely classify faults. Referring to Figure 2, a plasma processing chamber 202 is presented having an electrode subsystem including a lower electrode 204, an outer cover ring 206, and a lower cover ring 208. For ease of illustration, various other conventional subsystems are omitted, such as top electrodes, RF generators, turbo pumps, flow controllers, temperature controls, and the like.

A number of sensor instances 210, 212, and 214 are presented. For example, sensor 210 represents a pressure gauge that monitors chamber pressure. For example, sensor 212 represents a light radiation sensor that monitors the light emission of the plasma in the chamber. For example, sensor 214 represents a VI probe. Other sensors familiar to those skilled in the art can also be used.

The sensors 210, 212, and 214 provide sensor data to the fault detection/classification unit 220, which includes automatically performing fault detection based on the supplied sensor data and/or Or classified software and / or hardware. A fault database 222 representing a database of fault models (each of which includes fault condition fingerprints) is presented. Each fault model is a set of data that depicts a particular fault condition. In accordance with one or more embodiments of the present invention, these fault models are generated first (discussed later herein) and used to perform fault detection and/or fault classification.

The fault detection/classification unit 220 receives sensor data from a plurality of sensors. Since there are a large number of sensors in a typical modern plasma processing system, and each sensor transmits hundreds or more data samples per second, the fault detection/classification unit 220 performs an innovative data operation, such that It is possible to detect and/or classify faults efficiently and in a timely manner. This article will then discuss the operational aspects of the data that can detect and/or classify faults in a timely manner. The controller 224 utilizes fault identification and/or fault classification to control the plasma implement (for example, to stop the implement in time or to make an in-place adjustment to remove the fault before further damage occurs).

3 presents an example of a method of generating a fault model in accordance with an embodiment of the present invention. These fault models are stored in a fault database (e.g., fault database 222) for detecting and/or classifying faults during production. At step 302, sensor data for a plurality of wafers is collected from the sensor. The wafer associated with the sensor data item is labeled "bad", "good", or "unknown" (step 304). For example, an engineer knows that a particular wafer or wafer set is defective due to worn focus rings (whether intentional or not) and uses the sensor data associated with those wafers to derive a fault model of the wear focus ring. This excellent fault model is also quite useful for classification, just like any other fault model.

At step 306, the sensor channel is pre-filtered to exclude channels that exhibit minimal variation. Pre-filtered as part of the data manipulation to reduce the amount of data that needs to be processed during fault detection and/or analysis. Considering the large number of sensors involved and the fact that each sensor produces hundreds or more data samples per second, data manipulation is a useful step to improve the timeliness and efficiency of fault detection/filtering. In an embodiment, sensor channel data that does not contribute to fault detection (eg, no change to the fault of interest) is excluded during pre-filtering. Pre-filtering is discussed in more detail in the continuation flowchart.

At step 308, a weighted principal component analysis (PCA) is performed for the remaining data channels associated with all of the wafer data samples. The purpose of weighted PCA is to reduce the data dimension and to reduce the need for data processing to process another part of the data operation to improve the timeliness and efficiency of fault detection/filtering. The weighted PCA is discussed in more detail in the continuation flowchart. As a result of weighting the PCA, the data samples presenting the wafer are in the multi-dimensional PCA space (step 310).

At step 312, the fault signatures are partially defined using a "bad" sample indicated by the user. As described, the user can indicate that a particular sensor data set associated with a particular wafer is a known "bad" wafer sample and can use these data sets to construct a fault trace depicting the known fault feature. In general, the fault trace is a vector in the PCA space. The fault traces are discussed in more detail in the continuation diagram.

At step 314, the wafer-related data samples are presented in a 2-D correlation map. This diagram helps to efficiently analyze possible fault conditions based on data sample strength (fault severity) and similarity to fault traces (angle between sample data and fault traces). The schematic diagram in the 2-D correlation diagram will be discussed in more detail in the continuation diagram.

At step 316, the fault boundary is calculated from the 2-D correlation map (which defines the boundary of the data parameter that is considered to be faulty). The calculation of fault boundaries is discussed in more detail in the continuation diagram.

At step 318, the robustness of the fault model is verified. The fault model (as used herein) includes at least fault maps, fault boundaries, and PCA parameters (eg, PCA coefficients associated with PCA channels of the weighted PCA). The verification of the fault model is discussed in more detail in the continuation diagram.

4 presents steps associated with automatic channel filtering (pre-filtering step 306 of FIG. 3) in accordance with an embodiment of the present invention. The steps of Figure 4 represent an advantageous embodiment - by deleting channels that do not contribute to variation, there are other ways to perform pre-filtering to reduce the amount of data channels.

At step 402, summary statistics (eg, average, median, maximum, minimum, noise, etc.) are calculated for individual channels of all wafers submitted. In one embodiment, local linear matching is used to estimate the noise. In one embodiment, for example, if there are 10 wafers and 200 sensor data channels, as part of step 402, Summary statistics will be calculated for all 10 channels and all 200 channels of each wafer.

At step 404, the variance of the median statistic is calculated for each channel of all wafer samples. For example, if channel #37 measures pressure and 17mT is the median pressure reading for wafer #5 and 19mT is the median for wafer #6, median statistics are calculated for channel #37 for all wafers. The number of variations. For example, the variance number can be expressed as a standard deviation quantity. Even if some samples in the channel are missed from time to time during sensor data collection, the median tends to provide a more reliable statistic for channel data, so it is best to use the median. However, in some embodiments, other statistics (such as an average) are also used.

At step 406, the data channel deemed to be unchanged for the fault condition is excluded (i.e., the data change is not sufficiently significant to help identify the fault condition). An advantageous method of excluding a non-changing data channel involves comparing the variance of the aforementioned median statistic of the channel to a particular threshold, such as the noise threshold or channel resolution of the channel. The channel resolution depends on the sensor design characteristics and can be read from the pre-filled list of step 408. Other methods of excluding non-changing data channels can also be used.

As can be seen in step 406, if the median statistic of the data channel is zero, or if the median statistic of the data channel is less than a multiple of the noise, or if the data channel If the median statistic is less than the channel resolution (ie, the manufacturing tolerances associated with the channel due to the manufacture of the sensor, transmission line tolerances, etc.), the data channel is considered to be insufficiently variable. Cannot be included. In this case, the data channel is excluded (step 410). Otherwise, the data channel is incorporated to construct a fault model and/or to detect and/or analyze the fault (step 412).

FIG. 5 presents a weighted PCA method for reducing data dimensions in accordance with an embodiment of the present invention. In general, in PCA, the correlation between channel variances is analyzed to reduce the data dimension. For example, instead of presenting variances in a 100-dimensional data space, most of the variances are captured in a less-dimensional PCA space. In constructing the fault detection model, it is expected to capture the channel variation from the normal sample to the bad sample in the PCA data space.

However, if it is compared to the "good" and "unknown" sample sizes, the "bad" sample Very small amounts, the contribution of "bad" samples may be too negligible, making it difficult to capture the variance in the PCA data space.

In the weighted PCA method of FIG. 5, an acceptable threshold ratio of the "bad" sample to the "good" and "unknown" samples is predetermined. If the "bad" sample is below the ratio for the "good" and "unknown" samples, copy the "bad" sample until the threshold is reached. By doing so, even if the "bad" sample size is too small to capture the expected number of variations in the PCA data space, the expected variance can be captured between the "bad" sample and the good/unknown sample.

5, in step 502 expected to contain "bad" sample N ₁ and "good" and / or "unknown" group of data samples N _2. In step 504, to identify "bad" sample amount is less than N ₁ as "good" and / or a plurality of predetermined ratio "unknown" sample volume of N _2. The purpose of step 504 is to ascertain whether the amount of N ₁ in the bad sample is low enough to capture the expected number of variances contributed by the "bad" sample. Although any suitable ratio may be used, but in practice the ratio in FIG. 5 is set to N 1/10 ₂ is based.

If the "bad" samples is less than N _1, the replication "bad" samples until a sufficient phase having "poor" sample (step 506) to a predetermined ratio. Taking measurements to make sure there is enough "bad" ₁ sample size N (step 504) after, and then they execute in either the rate of PCA data set (step 508).

At step 510, maintaining a principal component analysis is sufficient to capture 80% (or several other desired percentages) of variance. Those familiar with PCA are quite clear about steps 508 and 510 and will not be described in detail herein.

6A, 6B, 6C, and 6D present steps for defining a fault vector (representing a faulty trace) while automatically addressing the chamber offset, in accordance with an embodiment of the present invention. In Figure 6A, all data samples are plotted in a reduced PCA data space (presented in Figure 3A as three-dimensional, but may have as many data dimensions as would be expected from the PCA step of Figure 5). Define the uncorrected fault vector V ₁ as a vector from the "good" (or worse, from "unknown") sample center toward the "bad" sample center.

The definition vector V ₀ is the chamber offset vector and is presented as shown in the reduced PCA data space of Figure 6A. The chamber shift over time affects the plasma (for example, it affects the "good" sample center over time) and needs to be taken into account to improve the accuracy of constructing the fault trace model. For example, if the "bad" sample is taken in different time segments than the "good" data sample, the chamber offset will represent the "bad" sample and the earlier "good" sample. A meaningful (non-trivial) contribution to the number of variations. In constructing the fault trace, more accurate fault detection and analysis can be achieved by separating the fault component from the chamber offset component. This is a significant improvement over the prior art.

The corrected fault vector V _F represents the necessary component of the uncorrected fault vector V ₁ after the solution of the chamber offset. In terms of mathematics, the operation of the corrected fault vector V _F is presented in Figure 6B. In FIG. 6B, the corrected fault vector V _F is equal to the unit vector (V ₀ ) of the uncorrected fault vector V ₁ minus the chamber offset multiplied by the unit vector V ₀ of the uncorrected fault vector V ₁ and the chamber offset. Pure volume product. The corrected fault vector V _F is seen in Figure 6A, which is placed along the line connecting the "bad" sample center to the offset vector V ₀ and perpendicular to the offset vector V ₀ . The calculation of the uncorrected fault vector V ₁ is presented in Figure 6C, while the calculation of the offset vector V ₀ is presented in Figure 6D. From the uncorrected fault vector V ₁ obtained using Fig. 6C and the calculation using the offset vector V ₀ obtained in Fig. 6D, the corrected fault vector can be obtained using the equation shown in Fig. 6B.

Referring to Figure 6C, it is determined at step 630 whether there is at least one sample labeled as good. In detail, the sample includes the set of data that characterizes the data channel for a single wafer. In the situation of Figure 6, the samples are in the PCA data space. If a good sample, defines the uncorrected fault vector V "bad" center of the sample (step 632) extending from "good" to _a center of the sample. Another way to define the uncorrected fault vector V ₁ extending to the center of the sample "bad" center of the sample (step 634) from the "unknown." In other words, if there are any "good" samples, use the "good" sample center to define the uncorrected fault vector V ₁ .

Referring complete offset vector V ₀ computed 6D, the method of the fit (e.g., the fit linear least squares) on all lines were "good" and "unknown" sample (step 652). If fitness (goodness-of-fit) below (step 654) a particular threshold (0.7, but can vary depending upon the desired in the example of FIG. 6D), displacement vector V ₀ is set to zero (step 656). On the other hand, if the fitness is higher than (step 654) threshold, a further check is performed to see if the data samples projected on this fit V ₀ have a good correlation with their time stamps (because expected As time passes, the data samples will be distributed along the offset vector in the direction of the offset).

Thus at step 658, to identify the relevance of their time stamp data as projected on the fit of the V _0. If the correlation exceeds another threshold (0.7, but may vary depending upon the desired in the example of FIG. 6D), the fit is accepted as an offset vector V ₀ (step 660). Otherwise, the offset vector V ₀ is set to zero (step 656).

Once the fault vector V _F is defined, the correlation between any wafer sample vector V _k and the fault vector V _F can be calculated to detect whether the wafer sample V _k exhibits a fault characteristic and, if so, to classify the fault ( By repeatedly comparing the wafer sample vector _Vk with different fault vectors representing different faults). 7A shows that the correlation can be quantified by the r(k) term, which represents the fault strength when the wafer sample vector _{Vk is} projected onto the fault vector _VF . This relationship is mathematically depicted in Figure 7B. Furthermore, the angle θ between the wafer sample vector _Vk and the fault vector _VF reflects how the wafer sample is similar to the fault. This relationship is mathematically depicted in Figure 7C.

Various samples can be calculated relative to the fault vector V _F to achieve presentation of the data samples in the 2-D coordinate system, and each sample k is plotted on r(k), θ(k).

Figure 8 illustrates the steps of setting a fault track boundary in accordance with an embodiment of the present invention. Once the data samples are drawn in the 2-D coordinate system (each sample is plotted on r(k), θ(k)), the boundary of the fault condition can be established. Figure 8 presents an example of a 2-D data sample map, with "good" or "unknown" samples gathered in region 802. The "bad" samples are clustered in area 804. To set the boundary of the fault condition represented by the "bad" sample in region 804, the maximum value θ(max) of the "bad" sample is identified, which exhibits the maximum angle θ among the "bad" samples.

In order to provide a marginal error, the angular boundary of the fault condition is a multiple of the maximum value θ(max) (integer or a multiple of a non-integer). In the example of Fig. 8, point 808 represents a sample having a maximum value θ(max) from the horizontal axis (θ = zero). The angle θ(max) associated with the point 808 sample is multiplied by 1.2 (an arbitrary number and may vary as desired) to obtain the maximum boundary value θ(b).

To provide a marginal error, the radius r associated with the data sample with the smallest r(k) is identified. The minimum radius r(k) is multiplied by a number of fractional constants (0.8 in the example of Figure 8 but Can vary with expectations). Referring to Figure 8, boundary 812 represents the boundary of the fault condition, and the data samples falling in boundary 812 can be classified into the fault condition.

If there is only a single "bad" data point, the θ(b) angle defining the fault condition additionally reflects the parameter values and some margins outside the "good" sample. In one embodiment, the minimum θ(min) of all good/unknown samples is reduced by a number of fractional constants to derive a θ(b) angle that defines the fault condition. Referring to Figure 8, it is assumed that point 814 represents a "good" sample with a minimum θ (min) value. Line 816 depicts the angle θ(min) associated with the smallest angle of the "good" or "unknown" sample. By reducing the minimum value θ(min) by a fractional constant (0.8 in this example, but the value can change as desired), the boundary angle θ(b) is obtained and is plotted by line 820 in FIG.

Figure 9 illustrates the steps of verifying a fault model in accordance with an embodiment of the present invention. At step 902, it is ascertained whether all of the "bad" samples are within the fault boundary and whether all of the "good" and "unknown" samples are outside the fault boundary. If the answer is no, the fault model is rejected (step 906). On the other hand, if both conditions of step 902 are true, it is further ascertained (step 904) whether each "bad" sample has been taken at least once from the "bad" sample cluster to re-verify the fault model. . If each "bad" sample has been taken at least once from the "bad" sample cluster to re-verify the failure model and the re-verification does not initiate rejection (step 902/906), then the model is accepted (step 908) .

On the other hand, if more than one "bad" sample is to be taken from the "bad" sample cluster to re-verify the failure model, then the "bad" sample that has not been taken out is taken from the bad sample cluster to facilitate Re-verification of the fault model (step 910). The fault vector and fault boundary calculations are performed again (at step 912) with the modified "bad" sample group (steps 312-316 of Figure 3). In the embodiment, each time a "bad" sample is taken from the "bad" sample cluster, the last "bad" sample taken is placed back into the cluster. In this way, only one "bad" sample is missing at a time to perform the recalculation of the fault vector and the fault boundary. In more than one embodiment, it may of course be possible to take a set of "bad" samples per turn (and replace the previous one). In this case, only one set of "bad" samples is missing at a time to perform a recalculation of the fault vector and the fault boundary.

Once all of the "bad" samples have been taken at least once and the fault model test is satisfactory, the verification is ended at step 908.

10 illustrates the steps of detecting and classifying more than one fault condition from wafer process data (block 1002) in accordance with an embodiment of the present invention. At step 1004, a summary statistic for the data channel is calculated. For example, this calculation has been discussed earlier with reference to FIG.

FIG. 10 also presents a fault model database (block 1006) that represents a pre-built data model 1006 for the fault model. The fault model contains fault traces, fault boundaries, and PCA parameters as discussed earlier. Different fault models depict features of different known faults (eg, worn edge rings, broken ground straps, incorrect chamber gaps, incorrect air pressure, etc.). A fault model from the fault model database is selected for testing (step 1008).

At step 1010, the wafer process sensor data associated with the wafer under test is converted to the same PCA space of the selected fault model of step 1008 to facilitate detection and classification for the selected fault model. At step 1012, the correlation (θ and r) of the wafer routing data is calculated for the fault vector V _F associated with the fault model selected at step 1008. At step 1014, it is ascertained whether the wafer process data under test is in the fault boundary of the fault model selected in step 1008. If the wafer process data under test is in the fault boundary of the selected fault model in step 1008, an alarm is initiated in step 1018 indicating that a possible fault consistent with the selected fault model of 1008 is detected and the fault pattern is classified.

On the other hand, if the wafer process data under test is not in the fault boundary of the selected fault model in step 1008, the program moves to step 1016 to find out if there are additional fault conditions to be tested again. In some cases, even if a possible failure has been detected, it may be desirable to test the wafer process data for other failure models to determine if there are multiple possible failures. If it is confirmed that there is another fault condition to test, the program moves to step 1008 to pick another fault model to test again. The program continues until it is confirmed in step 1016 that there are no additional fault conditions to test. At step 1020, a report is generated to report the fault correlation results and/or any possible faults that have been discovered/classified.

As can be seen from the foregoing, embodiments of the present invention facilitate automatic, efficient, and timely detection and classification of fault conditions despite the complex fault classification and bulky sensor data.

More importantly, embodiments of the present invention systematize the procedures for constructing fault models and lower the technical threshold. The operator no longer needs to know which data channel to include, how to resolve the baseline offset, and how to set the boundary conditions. These decisions have been systematized and automated in the fault model building process. In some cases, everything needed to construct a fault model is to identify the fault condition and mark the data samples associated with the wafer as appropriate "bad" labels.

The method of the present invention then automatically performs a data operation to pre-filter the data channel and further reduce the data dimension of the wafer material through the PCA. In particular, weighted PCA, automatic resolution of reference (chamber) offsets, and/or innovative fault classification methods based on fault models, 2-D data plotting and fault boundary settings make fault model construction procedures, fault detection and classification Be highly automated, robust and efficient.

Although the invention has been described in terms of several embodiments, variations, substitutions, and equivalents are possible within the scope of the invention. If a vocabulary "group" is used herein, the vocabulary is intended to have a generally understood mathematical meaning, covering zero, one, or more than one element. It is also noted that there are many alternative ways of implementing the methods and apparatus of embodiments of the present invention. Moreover, the utility of the embodiments of the present invention can be found in other applications. The Abstract section is provided for convenience and is intended to be read as it is limited by the number of words, and is not intended to limit the scope of the patent application. Accordingly, the scope of the following appended claims should be construed to cover all modifications, permutations, and equivalents in the true spirit and scope of the invention.

100‧‧‧lower electrode subsystem

102‧‧‧lower electrode

104‧‧‧External cover ring

106‧‧‧Top cover ring

108‧‧‧ wafer

200‧‧‧ various subcomponents of the plasma processing chamber

202‧‧‧The plasma processing chamber

204‧‧‧lower electrode

206‧‧‧External cover ring

208‧‧‧Top cover ring

210, 212, 214‧‧‧ sensors

220‧‧‧Fault detection/classification unit

222‧‧‧ Fault Database

224‧‧‧ Controller

302-318‧‧‧Steps

402-412‧‧‧Steps

502-510‧‧‧Steps

630-634‧‧ steps

652-660‧‧‧Steps

802, 804‧‧‧ sample collection area

808, 814‧‧ points

812‧‧‧ border

816, 820‧‧‧ straight line

902-912‧‧‧Steps

1002-1020‧‧‧Steps

The invention is illustrated by way of non-limiting example in the accompanying drawings, in which like reference numerals refer to the like elements, and wherein: Figure 1 shows an example of a lower electrode subsystem, including a lower electrode, external Cover ring, and top cover ring.

2 is a logic block diagram showing various sub-elements of a plasma processing chamber in accordance with an embodiment of the invention, which can automatically and timely detect fault conditions and automatically and timely classify faults.

3 presents an example of a method of generating a fault model in accordance with an embodiment of the present invention.

4 illustrates steps associated with automatic channel filtering in accordance with an embodiment of the present invention.

FIG. 5 presents a weighted PCA method for reducing data dimensions in accordance with an embodiment of the present invention.

6A, 6B, 6C, and 6D present a definition of a fault vector in accordance with an embodiment of the present invention. The table fault track is automatically incorporated into the chamber offset step.

7A, 7B, and 7C present correlation performance quantized by the r(k) term, and r(k) represents the fault strength when the wafer sample vector _Vk is projected onto the fault vector _VF .

Figure 8 illustrates the steps of setting a fault track boundary in accordance with an embodiment of the present invention.

Figure 9 illustrates the steps of verifying a fault model in accordance with an embodiment of the present invention.

10 illustrates the steps of detecting and/or classifying more than one fault condition from wafer process data in accordance with an embodiment of the present invention.

200. . . Various sub-components of the plasma processing chamber

202. . . Plasma processing chamber

204. . . Lower electrode

206. . . External cover ring

208. . . Top cover ring

210, 212, 214. . . Sensor

220. . . Fault detection/classification unit

222. . . Fault database

224. . . Controller

Claims (20)

A fault condition detecting method for detecting a first fault condition of a plurality of fault conditions during processing of the first substrate, the method comprising: collecting first data via a sensor during processing of the first substrate, wherein the first The data includes a first set of samples and a second set of samples; transmitting the first data to a faulty device; processing the first data via the faulty device, comprising: determining a corrected fault vector, comprising: determining an offset vector The offset vector corresponds to an offset of one or more parameters of a plasma processing chamber; determining an uncorrected fault vector based on the first set of samples and the second set of samples; and correcting according to the offset vector The uncorrected fault vector is generated to generate the corrected fault vector; and comparing the first portion of the first data with the plurality of fault models stored in a fault database, wherein the plurality of fault models respectively correspond to the number Fault conditions, wherein the plurality of fault models include the following: (i) a plurality of fault vectors, (ii) a plurality of fault boundaries, and (iii) an array principal component analysis (PCA) parameter, and wherein the plurality Impaired vector containing the fault has been corrected vector; and comparing the number of fault model based on the first part of the first data, the first detection of the fault condition. The method for detecting a fault condition according to claim 1, wherein the processing of the first data by the faulty device comprises: calculating a summary statistic value for each channel of the sensor; and pre-filtering the first Data to exclude a second portion of the first data from the first set of channels of the sensor, wherein the first set of channels has no change for the plurality of fault conditions. For example, the fault condition detection method of claim 2, wherein: The processing of the first data includes performing a weighted PCA on the first portion of the first data from the second set of channels of the sensor; and the first of the first data from the second set of channels The portion does not include the second portion of the first data from the first set of channels. The fault condition detecting method of claim 3, wherein: the processing of the first data comprises generating a set of data samples represented in a multi-dimensional PCA space according to one of the set of PCA parameters; The shift vector is based on the set of data samples present in the multi-dimensional PCA space; and the offset of the plasma processing chamber is represented by the offset vector in the multi-dimensional PCA space. The method for detecting a fault condition according to claim 4, wherein the comparison between the first part of the first data and the plurality of fault models comprises: a fault vector of the first fault model according to the plurality of fault models A correlation value of the first data is determined. The method for detecting a fault condition according to claim 5, wherein: the comparing the first part of the first data with the plurality of fault models comprises determining whether the first part of the first data is in the first a fault boundary of a fault model; if the first portion of the first data is in the fault boundary of the first fault model, generating an alert signal indicating a possible fault condition; and if the first data is The first portion identifies the second fault model of the one of the plurality of fault models and compares the first portion of the first data with the second fault model. The fault condition detecting method of claim 1, wherein each of the plurality of fault models is generated by collecting data sets from the sensor for a plurality of substrates, wherein the data is collected. The group includes the first data, and wherein the plurality of substrates comprise the first substrate; the title is respectively assigned to the data group, wherein each of the titles is identified as a set of bad materials, a set of good materials, or a group Unknown data; and pre-filtering the plurality of channels of the sensor to exclude data from the first portion of the plurality of channels, wherein the first portion of the plurality of channels does not contribute to fault detection, wherein the advance Filtering includes calculating a statistic summary for the second portion of the plurality of channels for each of the plurality of substrates. The fault condition detecting method of claim 7, wherein each of the titles is based on a component in the plasma processing chamber and a condition of one of the plurality of substrates. For example, in the fault condition detection method of claim 7, wherein calculating the statistical value summary includes using local linear adaptation to the data set. For example, the method for detecting a fault condition in claim 7 of the patent scope, wherein the pre-filtering further comprises calculating a variance in the summary of the statistics. A detection system for detecting and classifying a first fault condition of a plurality of fault conditions during processing of a first substrate in a plasma processing chamber, the system comprising: a sensor configured to Monitoring the plasma processing chamber during the processing of the first substrate; a fault database configured to store a plurality of fault models, wherein each of the plurality of fault models corresponds to the plurality of fault conditions; And a faulty device configured to: (i) collect first data from the sensor, wherein the first data includes a first set of samples and a second set of samples; and (ii) analyze the first The first part of the data includes: determining a corrected fault vector, comprising: determining an offset vector, the offset vector corresponding to the plasma processing chamber An offset of one or more parameters; determining an uncorrected fault vector based on the first set of samples and the second set of samples; and correcting the uncorrected fault vector based on the offset vector to generate the corrected fault vector; Comparing the first portion of the first data with at least one of the plurality of fault models, wherein the plurality of fault models include the following individuals (i) a plurality of fault vectors, (ii) a plurality of fault boundaries, and Iii) an array principal component analysis (PCA) parameter, wherein the plurality of fault vectors comprise the corrected fault vector, and wherein the at least one of the plurality of fault models includes the corrected fault vector; and based on the first data The comparison of the first portion with the at least one of the plurality of fault models detects the first fault condition. The detection system of claim 11 further includes a controller configured to control the detection system based on an output of the faulty device. The detection system of claim 11, wherein the faulty device is configured to: calculate a summary statistic value for each channel of the sensor; and pre-filter the first data to exclude from the sensor The second portion of the first data of the first group of channels, wherein the first group of channels has no change for the plurality of fault conditions. The detection system of claim 13 wherein: the faulty device is configured to perform a weighted PCA on the first portion of the first data from the second set of channels of the sensor; and from the The first portion of the first data of the second set of channels does not include the second portion of the first data from the first set of channels. For example, in the detection system of claim 14 of the patent scope, wherein: The faulty device is configured to generate a set of data samples represented in a multi-dimensional PCA space according to one of the array PCA parameters; the offset vector is based on the set of data samples represented in the multi-dimensional PCA space; The offset of the plasma processing chamber is represented by the offset vector in the multi-dimensional PCA space. The detection system of claim 15, wherein the faulty device is configured to determine a correlation value of the first data according to a fault vector of the first fault model of the plurality of fault models. The detecting system of claim 16 wherein the faulty device is configured to: compare the first portion of the first data with the plurality of fault models; and determine the first portion of the first data Whether it is in a fault boundary of the first fault model; if the first portion of the first data is in the fault boundary of the first fault model, generating an alarm signal indicating a possible fault condition; and if The first part of the first data identifies the second fault model of the one of the plurality of fault models outside the fault boundary of the first fault model, and compares the first portion of the first data with the The second failure model. The detection system of claim 11, wherein the faulty device is configured to generate each of the plurality of fault models by: collecting data sets from the sensor for each of the plurality of substrates, wherein The data set includes the first data, and wherein the plurality of substrates comprise the first substrate; the title is respectively assigned to the data group, wherein each of the titles is identified as a set of bad materials, a set of good materials, or a set of unknown data; and pre-filtering the plurality of channels of the sensor to exclude data from the first portion of the plurality of channels of the sensor, wherein the first portion of the plurality of channels does not contribute to Therefore Barrier detection, wherein the pre-filtering comprises calculating a statistic summary of the second portion of the plurality of channels for each of the plurality of substrates. The detection system of claim 18, wherein each of the titles is based on a component in the plasma processing chamber and a condition of one of the plurality of substrates. The detection system of claim 18, wherein the faulty device is configured to use local linear adaptation to the data set when calculating the statistical value summary.